refine oh-my-frontend orchestration

[fe-dudu]

Summary

• Simplify oh-my-frontend into risk-based frontend triage and orchestration.
• Keep Level 0 and bounded work lightweight; narrow task packets, specialist workers, review workers, routing, Mermaid updates, and durable documentation changes.
• Separate implementation guidance into frontend-engineering and keep worker reports evidence-oriented and unified.
• Harden the static eval grader with explicit skill selection, scoped IDs, complete workspace validation, failure propagation, and focused tests.
• Align plugin metadata at 0.2.0.

Why

The previous workflow could add planning, documentation, routing, or workers too broadly. The grader could also skip empty outputs and accept incomplete eval workspaces. This change makes orchestration proportional to risk and makes eval failures visible.

Validation

• go test -count=1 ./... in lint/
• go test -count=1 ./... in evals/
• go vet ./... in lint/ and evals/
• Skill linter: 51 file(s) checked, 0 diagnostic(s) found
• Taskfile/JSON/YAML parsing and git diff --check

The Korean mirror under docs/oh-my-frontend.ko/ is intentionally gitignored and is synchronized locally, but is not included in this PR.